Reducing Failure-Inducing Inputs

By construction, fuzzers create inputs that may be hard to read. This causes issues during debugging, when a human has to analyze the exact cause of the failure. In this chapter, we present techniques that automatically reduce and simplify failure-inducing inputs to a minimum in order to ease debugging.

from bookutils import YouTubeVideo
YouTubeVideo('JOv1xGVdXAU')

Prerequisites

  • The simple "delta debugging" technique for reduction has no specific prerequisites.
  • As reduction is typically used together with fuzzing, reading the chapter on basic fuzzing is a good idea.
  • The later grammar-based techniques require knowledge on derivation trees and parsing.

Synopsis

To use the code provided in this chapter, write

>>> from fuzzingbook.Reducer import <identifier>

and then make use of the following features.

A reducer takes a failure-inducing input and reduces it to the minimum that still reproduces the failure. This chapter provides Reducer classes that implement such reducers.

Here is a simple example: An arithmetic expression causes an error in the Python interpreter:

>>> !python -c 'x = 1 + 2 * 3 / 0'
Traceback (most recent call last):
  File "<string>", line 1, in <module>
ZeroDivisionError: division by zero

Can we reduce this input to a minimum? To use a Reducer, one first has to build a Runner whose outcome is FAIL if the precise error occurs. We therefore build a ZeroDivisionRunner whose run() method will specifically return a FAIL outcome if a ZeroDivisionError occurs.

>>> from Fuzzer import ProgramRunner
>>> import subprocess
>>> class ZeroDivisionRunner(ProgramRunner):
>>>     """Make outcome 'FAIL' if ZeroDivisionError occurs"""
>>> 
>>>     def run(self, inp: str = "") -> Tuple[subprocess.CompletedProcess, Outcome]:
>>>         process, outcome = super().run(inp)
>>>         if process.stderr.find('ZeroDivisionError') >= 0:
>>>             outcome = 'FAIL'
>>>         return process, outcome

If we feed this expression into a ZeroDivisionRunner, it will produce an outcome of FAIL as designed.

>>> python_input = "x = 1 + 2 * 3 / 0"
>>> python_runner = ZeroDivisionRunner("python")
>>> process, outcome = python_runner.run(python_input)
>>> outcome
'FAIL'

Delta Debugging is a simple and robust reduction algorithm. We can tie a DeltaDebuggingReducer to this runner, and have it determine the substring that causes the python program to fail:

>>> dd = DeltaDebuggingReducer(python_runner)
>>> dd.reduce(python_input)
'3/0'

The input is reduced to the minimum: We get the essence of the division by zero.

DeltaDebuggingReducer DeltaDebuggingReducer reduce() CachingReducer CachingReducer reset() test() DeltaDebuggingReducer->CachingReducer Reducer Reducer __init__() reduce() reset() test() CachingReducer->Reducer GrammarReducer GrammarReducer __init__() reduce() alternate_reductions() parse() reduce_subtree() reduce_tree() subtrees_with_symbol() symbol_reductions() GrammarReducer->CachingReducer Legend Legend •  public_method() •  private_method() •  overloaded_method() Hover over names to see doc

Why Reducing?

At this point, we have seen a number of test generation techniques that all in some form produce inputs in order to trigger failures. If they are successful – that is, the program actually fails – we must find out why the failure occurred and how to fix it.

Here's an example of such a situation. We have a class MysteryRunner with a run() method that – given its code – can occasionally fail. But under which circumstances does this actually happen? We have deliberately obscured the exact condition in order to make this non-obvious.

from bookutils import quiz
from typing import Tuple, List, Sequence, Any, Optional
from ExpectError import ExpectError
from Fuzzer import RandomFuzzer, Runner, Outcome
import re
class MysteryRunner(Runner):
    def run(self, inp: str) -> Tuple[str, Outcome]:
        x = inp.find(chr(0o17 + 0o31))
        y = inp.find(chr(0o27 + 0o22))
        if x >= 0 and y >= 0 and x < y:
            return (inp, Runner.FAIL)
        else:
            return (inp, Runner.PASS)

Let us fuzz the function until we find a failing input.

mystery = MysteryRunner()
random_fuzzer = RandomFuzzer()
while True:
    inp = random_fuzzer.fuzz()
    result, outcome = mystery.run(inp)
    if outcome == mystery.FAIL:
        break
failing_input = result
failing_input
' 7:,>((/$$-/->.;.=;(.%!:50#7*8=$&&=$9!%6(4=&69\':\'<3+0-3.24#7=!&60)2/+";+<7+1<2!4$>92+$1<(3%&5\'\'>#'

Something in this input causes MysteryRunner to fail. But what is it?

Manual Input Reduction

One important step in the debugging process is reduction – that is, to identify those circumstances of a failure that are relevant for the failure to occur, and to omit (if possible) those parts that are not. As Kernighan and Pike [Kernighan et al, 1999] put it:

For every circumstance of the problem, check whether it is relevant for the problem to occur. If it is not, remove it from the problem report or the test case in question.

Specifically for inputs, they suggest a divide and conquer process:

Proceed by binary search. Throw away half the input and see if the output is still wrong; if not, go back to the previous state and discard the other half of the input.

This is something we can easily try out, using our last generated input:

failing_input
' 7:,>((/$$-/->.;.=;(.%!:50#7*8=$&&=$9!%6(4=&69\':\'<3+0-3.24#7=!&60)2/+";+<7+1<2!4$>92+$1<(3%&5\'\'>#'

For instance, we can see whether the error still occurs if we only feed in the first half:

half_length = len(failing_input) // 2   # // is integer division
first_half = failing_input[:half_length]
mystery.run(first_half)
(" 7:,>((/$$-/->.;.=;(.%!:50#7*8=$&&=$9!%6(4=&69':", 'PASS')

Nope – the first half alone does not suffice. Maybe the second half?

second_half = failing_input[half_length:]
mystery.run(second_half)
('\'<3+0-3.24#7=!&60)2/+";+<7+1<2!4$>92+$1<(3%&5\'\'>#', 'PASS')

This did not go so well either. We may still proceed by cutting away smaller chunks – say, one character after another. If our test is deterministic and easily repeated, it is clear that this process eventually will yield a reduced input. But still, it is a rather inefficient process, especially for long inputs. What we need is a strategy that effectively minimizes a failure-inducing input – a strategy that can be automated.

Delta Debugging

One strategy to effectively reduce failure-inducing inputs is delta debugging [Zeller et al, 2002]. Delta Debugging implements the "binary search" strategy, as listed above, but with a twist: If neither half fails (also as above), it keeps on cutting away smaller and smaller chunks from the input, until it eliminates individual characters. Thus, after cutting away the first half, we cut away the first quarter, the second quarter, and so on.

Let us illustrate this on our example, and see what happens if we cut away the first quarter.

quarter_length = len(failing_input) // 4
input_without_first_quarter = failing_input[quarter_length:]
mystery.run(input_without_first_quarter)
('50#7*8=$&&=$9!%6(4=&69\':\'<3+0-3.24#7=!&60)2/+";+<7+1<2!4$>92+$1<(3%&5\'\'>#',
 'FAIL')

Ah! This has failed, and reduced our failing input by 25%. Let's remove another quarter.

input_without_first_and_second_quarter = failing_input[quarter_length * 2:]
mystery.run(input_without_first_and_second_quarter)
('\'<3+0-3.24#7=!&60)2/+";+<7+1<2!4$>92+$1<(3%&5\'\'>#', 'PASS')

This is not too surprising, as we had that one before:

second_half
'\'<3+0-3.24#7=!&60)2/+";+<7+1<2!4$>92+$1<(3%&5\'\'>#'
input_without_first_and_second_quarter
'\'<3+0-3.24#7=!&60)2/+";+<7+1<2!4$>92+$1<(3%&5\'\'>#'

How about removing the third quarter, then?

input_without_first_and_third_quarter = failing_input[quarter_length:
                                                      quarter_length * 2] + failing_input[quarter_length * 3:]
mystery.run(input_without_first_and_third_quarter)
("50#7*8=$&&=$9!%6(4=&69':<7+1<2!4$>92+$1<(3%&5''>#", 'PASS')

Ok. Let us remove the fourth quarter.

input_without_first_and_fourth_quarter = failing_input[quarter_length:quarter_length * 3]
mystery.run(input_without_first_and_fourth_quarter)
('50#7*8=$&&=$9!%6(4=&69\':\'<3+0-3.24#7=!&60)2/+";+', 'FAIL')

Yes! This has succeeded. Our input is now 50% smaller.

We have now tried to remove pieces that make up $\frac{1}{2}$ and $\frac{1}{4}$ of the original failing string. In the next iteration, we would go and remove even smaller pieces – $\frac{1}{8}$, $\frac{1}{16}$ and so on. We continue until we are down to $\frac{1}{97}$ – that is, individual characters.

However, this is something we happily let a computer do for us. We first introduce a Reducer class as an abstract superclass for all kinds of reducers. The test() method runs a single test (with logging, if wanted); the reduce() method will eventually reduce an input to the minimum.

class Reducer:
    """Base class for reducers."""

    def __init__(self, runner: Runner, log_test: bool = False) -> None:
        """Attach reducer to the given `runner`"""
        self.runner = runner
        self.log_test = log_test
        self.reset()

    def reset(self) -> None:
        """Reset the test counter to zero. To be extended in subclasses."""
        self.tests = 0

    def test(self, inp: str) -> Outcome:
        """Test with input `inp`. Return outcome.
        To be extended in subclasses."""

        result, outcome = self.runner.run(inp)
        self.tests += 1
        if self.log_test:
            print("Test #%d" % self.tests, repr(inp), repr(len(inp)), outcome)
        return outcome

    def reduce(self, inp: str) -> str:
        """Reduce input `inp`. Return reduced input.
        To be defined in subclasses."""

        self.reset()
        # Default: Don't reduce
        return inp

The CachingReducer variant saves test results, such that we don't have to run the same tests again and again:

class CachingReducer(Reducer):
    """A reducer that also caches test outcomes"""

    def reset(self):
        super().reset()
        self.cache = {}

    def test(self, inp):
        if inp in self.cache:
            return self.cache[inp]

        outcome = super().test(inp)
        self.cache[inp] = outcome
        return outcome

Here comes the Delta Debugging reducer. Delta Debugging implements the strategy sketched above: It first removes larger chunks of size $\frac{1}{2}$; if this does not fail, then we proceed to chunks of size $\frac{1}{4}$, then $\frac{1}{8}$ and so on.

Our implementation uses almost the same Python code as Zeller in [Zeller et al, 2002]; the only difference is that it has been adapted to work on Python 3 and our Runner framework. The variable n (initially 2) indicates the granularity – in each step, chunks of size $\frac{1}{n}$ are cut away. If none of the test fails (some_complement_is_failing is False), then n is doubled – until it reaches the length of the input.

class DeltaDebuggingReducer(CachingReducer):
    """Reduce inputs using delta debugging."""

    def reduce(self, inp: str) -> str:
        """Reduce input `inp` using delta debugging. Return reduced input."""

        self.reset()
        assert self.test(inp) != Runner.PASS

        n = 2     # Initial granularity
        while len(inp) >= 2:
            start = 0.0
            subset_length = len(inp) / n
            some_complement_is_failing = False

            while start < len(inp):
                complement = inp[:int(start)] + \
                    inp[int(start + subset_length):]

                if self.test(complement) == Runner.FAIL:
                    inp = complement
                    n = max(n - 1, 2)
                    some_complement_is_failing = True
                    break

                start += subset_length

            if not some_complement_is_failing:
                if n == len(inp):
                    break
                n = min(n * 2, len(inp))

        return inp

To see how the DeltaDebuggingReducer works, let us run it on our failing input. With each step, we see how the remaining input gets smaller and smaller, until only two characters remain:

dd_reducer = DeltaDebuggingReducer(mystery, log_test=True)
dd_reducer.reduce(failing_input)
Test #1 ' 7:,>((/$$-/->.;.=;(.%!:50#7*8=$&&=$9!%6(4=&69\':\'<3+0-3.24#7=!&60)2/+";+<7+1<2!4$>92+$1<(3%&5\'\'>#' 97 FAIL
Test #2 '\'<3+0-3.24#7=!&60)2/+";+<7+1<2!4$>92+$1<(3%&5\'\'>#' 49 PASS
Test #3 " 7:,>((/$$-/->.;.=;(.%!:50#7*8=$&&=$9!%6(4=&69':" 48 PASS
Test #4 '50#7*8=$&&=$9!%6(4=&69\':\'<3+0-3.24#7=!&60)2/+";+<7+1<2!4$>92+$1<(3%&5\'\'>#' 73 FAIL
Test #5 "50#7*8=$&&=$9!%6(4=&69':<7+1<2!4$>92+$1<(3%&5''>#" 49 PASS
Test #6 '50#7*8=$&&=$9!%6(4=&69\':\'<3+0-3.24#7=!&60)2/+";+' 48 FAIL
Test #7 '\'<3+0-3.24#7=!&60)2/+";+' 24 PASS
Test #8 "50#7*8=$&&=$9!%6(4=&69':" 24 PASS
Test #9 '9!%6(4=&69\':\'<3+0-3.24#7=!&60)2/+";+' 36 FAIL
Test #10 '9!%6(4=&69\':=!&60)2/+";+' 24 FAIL
Test #11 '=!&60)2/+";+' 12 PASS
Test #12 "9!%6(4=&69':" 12 PASS
Test #13 '=&69\':=!&60)2/+";+' 18 PASS
Test #14 '9!%6(4=!&60)2/+";+' 18 FAIL
Test #15 '9!%6(42/+";+' 12 PASS
Test #16 '9!%6(4=!&60)' 12 FAIL
Test #17 '=!&60)' 6 PASS
Test #18 '9!%6(4' 6 PASS
Test #19 '6(4=!&60)' 9 FAIL
Test #20 '6(460)' 6 FAIL
Test #21 '60)' 3 PASS
Test #22 '6(4' 3 PASS
Test #23 '(460)' 5 FAIL
Test #24 '460)' 4 PASS
Test #25 '(0)' 3 FAIL
Test #26 '0)' 2 PASS
Test #27 '(' 1 PASS
Test #28 '()' 2 FAIL
Test #29 ')' 1 PASS
'()'

Now we know why MysteryRunner fails – it suffices that the input contains two matching parentheses. Delta Debugging determines this in 29 steps. Its result is 1-minimal, meaning that every character contained is required to produce the error; removing any (as seen in tests #27 and #29, above) no longer makes the test fail. This property is guaranteed by the delta debugging algorithm, which in its last stage always tries to delete characters one by one.

A reduced test case such as the one above has many advantages:

  • A reduced test case reduces the cognitive load of the programmer. The test case is shorter and focused, and thus does not burden the programmer with irrelevant details. A reduced input typically leads to shorter executions and smaller program states, both of which reduce the search space as it comes to understanding the bug. In our case, we have eliminated lots of irrelevant input – only the two characters the reduced input contains are relevant.

  • A reduced test case is easier to communicate. All one needs here is the summary: MysteryRunner fails on "()", which is much better than MysteryRunner fails on a 4100-character input (attached).

  • A reduced test case helps in identifying duplicates. If similar bugs have been reported already, and all of them have been reduced to the same cause (namely that the input contains matching parentheses), then it becomes obvious that all these bugs are different symptoms of the same underlying cause – and would all be resolved at once with one code fix.

How effective is delta debugging? In the best case (when the left half or the right half fails), the number of tests is logarithmic proportional to the length $n$ of an input (i.e., $O(\log_2 n)$); this is the same complexity as binary search. In the worst case, though, delta debugging can require a number of tests proportional to $n^2$ (i.e., $O(n^2)$) – this happens in the case when we are down to character granularity, and we have to repeatedly tried to delete all characters, only to find that deleting the last character results in a failure [Zeller et al, 2002]. (This is a pretty pathological situation, though.)

In general, delta debugging is a robust algorithm that is easy to implement, easy to deploy, and easy to use – provided that the underlying test case is deterministic and runs quickly enough to warrant a number of experiments. As these are the same prerequisites that make fuzzing effective, delta debugging makes an excellent companion to fuzzing.

Quiz

What happens if the function under test does not fail?





Indeed, DeltaDebugger checks if its assumptions hold. If not, an assertion fails.

with ExpectError():
    dd_reducer.reduce("I am a passing input")
Test #1 'I am a passing input' 20 PASS
Traceback (most recent call last):
  File "/var/folders/n2/xd9445p97rb3xh7m1dfx8_4h0006ts/T/ipykernel_77032/3080932113.py", line 2, in <cell line: 1>
    dd_reducer.reduce("I am a passing input")
  File "/var/folders/n2/xd9445p97rb3xh7m1dfx8_4h0006ts/T/ipykernel_77032/2870451724.py", line 8, in reduce
    assert self.test(inp) != Runner.PASS
AssertionError (expected)

Grammar-Based Input Reduction

If the input language is syntactically complex, delta debugging may take several attempts at reduction, and may not be able to reduce inputs at all. In the second half of this chapter, we thus introduce an algorithm named Grammar-Based Reduction (or GRABR for short) that makes use of grammars to reduce syntactically complex inputs.

Lexical Reduction vs. Syntactic Rules

Despite its general robustness, there are situations in which delta debugging might be inefficient or outright fail. As an example, consider some expression input such as 1 + (2 * 3). Delta debugging requires a number of tests to simplify the failure-inducing input, but it eventually returns a minimal input

expr_input = "1 + (2 * 3)"
dd_reducer = DeltaDebuggingReducer(mystery, log_test=True)
dd_reducer.reduce(expr_input)
Test #1 '1 + (2 * 3)' 11 FAIL
Test #2 '2 * 3)' 6 PASS
Test #3 '1 + (' 5 PASS
Test #4 '+ (2 * 3)' 9 FAIL
Test #5 '+ ( 3)' 6 FAIL
Test #6 ' 3)' 3 PASS
Test #7 '+ (' 3 PASS
Test #8 ' ( 3)' 5 FAIL
Test #9 '( 3)' 4 FAIL
Test #10 '3)' 2 PASS
Test #11 '( ' 2 PASS
Test #12 '(3)' 3 FAIL
Test #13 '()' 2 FAIL
Test #14 ')' 1 PASS
Test #15 '(' 1 PASS
'()'

Looking at the tests, above, though, only few of them actually represent syntactically valid arithmetic expressions. In a practical setting, we may want to test a program which actually parses such expressions, and which would reject all invalid inputs. We define a class EvalMysteryRunner which first parses the given input (according to the rules of our expression grammar), and only if it fits would it be passed to our original MysteryRunner. This simulates a setting in which we test an expression interpreter, and in which only valid inputs can trigger the bug.

from Grammars import EXPR_GRAMMAR
from Parser import EarleyParser, Parser  # minor dependency
class EvalMysteryRunner(MysteryRunner):
    def __init__(self) -> None:
        self.parser = EarleyParser(EXPR_GRAMMAR)

    def run(self, inp: str) -> Tuple[str, Outcome]:
        try:
            tree, *_ = self.parser.parse(inp)
        except SyntaxError:
            return (inp, Runner.UNRESOLVED)

        return super().run(inp)
eval_mystery = EvalMysteryRunner()

Under these circumstances, it turns out that delta debugging utterly fails. None of the reductions it applies yield a syntactically valid input, so the input as a whole remains as complex as it was before.

dd_reducer = DeltaDebuggingReducer(eval_mystery, log_test=True)
dd_reducer.reduce(expr_input)
Test #1 '1 + (2 * 3)' 11 FAIL
Test #2 '2 * 3)' 6 UNRESOLVED
Test #3 '1 + (' 5 UNRESOLVED
Test #4 '+ (2 * 3)' 9 UNRESOLVED
Test #5 '1 2 * 3)' 8 UNRESOLVED
Test #6 '1 + ( 3)' 8 UNRESOLVED
Test #7 '1 + (2 *' 8 UNRESOLVED
Test #8 ' + (2 * 3)' 10 UNRESOLVED
Test #9 '1+ (2 * 3)' 10 UNRESOLVED
Test #10 '1 (2 * 3)' 9 UNRESOLVED
Test #11 '1 + 2 * 3)' 10 UNRESOLVED
Test #12 '1 + ( * 3)' 10 UNRESOLVED
Test #13 '1 + (2 3)' 9 UNRESOLVED
Test #14 '1 + (2 *3)' 10 UNRESOLVED
Test #15 '1 + (2 * ' 9 UNRESOLVED
Test #16 '1  (2 * 3)' 10 UNRESOLVED
Test #17 '1 +(2 * 3)' 10 UNRESOLVED
Test #18 '1 + (2* 3)' 10 UNRESOLVED
Test #19 '1 + (2  3)' 10 UNRESOLVED
Test #20 '1 + (2 * )' 10 UNRESOLVED
Test #21 '1 + (2 * 3' 10 UNRESOLVED
'1 + (2 * 3)'

This behavior is possible if the program under test has several constraints regarding input validity. Delta debugging is not aware of these constraints (nor of the input structure in general), so it might violate these constraints again and again.

A Grammar-Based Reduction Approach

To reduce inputs with high syntactical complexity, we use another approach: Rather than reducing the input string, we reduce the tree representing its structure. The general idea is to start with a derivation tree coming from parsing the input, and then substitute subtrees by smaller subtrees of the same type. These alternate subtrees can either come

  1. From the tree itself, or
  2. By applying an alternate grammar expansion using elements from the tree.

Let us show these two strategies using an example. We start with a derivation tree from an arithmetic expression:

from Grammars import Grammar
from GrammarFuzzer import all_terminals, expansion_to_children, display_tree
derivation_tree, *_ = EarleyParser(EXPR_GRAMMAR).parse(expr_input)
display_tree(derivation_tree)
0 <start> 1 <expr> 0->1 2 <term> 1->2 7 + 1->7 8 <expr> 1->8 3 <factor> 2->3 4 <integer> 3->4 5 <digit> 4->5 6 1 (49) 5->6 9 <term> 8->9 10 <factor> 9->10 11 ( (40) 10->11 12 <expr> 10->12 24 ) (41) 10->24 13 <term> 12->13 14 <factor> 13->14 18 * 13->18 19 <term> 13->19 15 <integer> 14->15 16 <digit> 15->16 17 2 (50) 16->17 20 <factor> 19->20 21 <integer> 20->21 22 <digit> 21->22 23 3 (51) 22->23

Simplifying by Replacing Subtrees

To simplify this tree, we could replace any <expr> symbol up in the tree with some <expr> subtree down in the tree. For instance, we could replace the uppermost <expr> with its right <expr> subtree, yielding the string (2 + 3):

import copy
new_derivation_tree = copy.deepcopy(derivation_tree)
# We really should have some query language
sub_expr_tree = new_derivation_tree[1][0][1][2]
display_tree(sub_expr_tree)
0 <expr> 1 <term> 0->1 2 <factor> 1->2 3 ( (40) 2->3 4 <expr> 2->4 16 ) (41) 2->16 5 <term> 4->5 6 <factor> 5->6 10 * 5->10 11 <term> 5->11 7 <integer> 6->7 8 <digit> 7->8 9 2 (50) 8->9 12 <factor> 11->12 13 <integer> 12->13 14 <digit> 13->14 15 3 (51) 14->15
new_derivation_tree[1][0] = sub_expr_tree
display_tree(new_derivation_tree)
0 <start> 1 <expr> 0->1 2 <term> 1->2 3 <factor> 2->3 4 ( (40) 3->4 5 <expr> 3->5 17 ) (41) 3->17 6 <term> 5->6 7 <factor> 6->7 11 * 6->11 12 <term> 6->12 8 <integer> 7->8 9 <digit> 8->9 10 2 (50) 9->10 13 <factor> 12->13 14 <integer> 13->14 15 <digit> 14->15 16 3 (51) 15->16
all_terminals(new_derivation_tree)
'(2 * 3)'

Replacing one subtree by another only works as long as individual elements such as <expr> occur multiple times in our tree. In the reduced new_derivation_tree, above, we could replace further <expr> trees only once more.

Simplifying by Alternative Expansions

A second means to simplify this tree is to apply alternative expansions. That is, for a symbol, we check whether there is an alternative expansion with a smaller number of children. Then, we replace the symbol with the alternative expansion, filling in needed symbols from the tree.

As an example, consider the new_derivation_tree above. The applied expansion for <term> has been

<term> ::= <term> * <factor>

Let us replace this with the alternative expansion:

<term> ::= <factor>
term_tree = new_derivation_tree[1][0][1][0][1][0][1][1][1][0]
display_tree(term_tree)
0 <term> 1 <factor> 0->1 5 * 0->5 6 <term> 0->6 2 <integer> 1->2 3 <digit> 2->3 4 2 (50) 3->4 7 <factor> 6->7 8 <integer> 7->8 9 <digit> 8->9 10 3 (51) 9->10
shorter_term_tree = term_tree[1][2]
display_tree(shorter_term_tree)
0 <term> 1 <factor> 0->1 2 <integer> 1->2 3 <digit> 2->3 4 3 (51) 3->4
new_derivation_tree[1][0][1][0][1][0][1][1][1][0] = shorter_term_tree
display_tree(new_derivation_tree)
0 <start> 1 <expr> 0->1 2 <term> 1->2 3 <factor> 2->3 4 ( (40) 3->4 5 <expr> 3->5 11 ) (41) 3->11 6 <term> 5->6 7 <factor> 6->7 8 <integer> 7->8 9 <digit> 8->9 10 3 (51) 9->10
all_terminals(new_derivation_tree)
'(3)'

If we replace derivation subtrees by (smaller) subtrees, and if we search for alternate expansions that again yield smaller subtrees, we can systematically simplify the input. This could be much faster than delta debugging, as our inputs would always be syntactically valid. However, we need a strategy for when to apply which simplification rule. This is what we develop in the remainder of this section.

A Class for Reducing with Grammars

We introduce the GrammarReducer class, which is again a Reducer. Note that we derive from CachingReducer, as the strategy will produce several duplicates.

class GrammarReducer(CachingReducer):
    """Reduce inputs using grammars"""

    def __init__(self, runner: Runner, parser: Parser, *,
                 log_test: bool = False, log_reduce: bool = False):
        """Constructor.
        `runner` is the runner to be used.
        `parser` is the parser to be used.
        `log_test` - if set, show tests and results.
        `log_reduce` - if set, show reduction steps.
        """

        super().__init__(runner, log_test=log_test)
        self.parser = parser
        self.grammar = parser.grammar()
        self.start_symbol = parser.start_symbol()
        self.log_reduce = log_reduce
        self.try_all_combinations = False

A Few Helpers

We define a number of helper functions, which we will need for our strategy. tree_list_to_string() does what the name suggest, creating a string from a list of derivation trees:

from GrammarFuzzer import DerivationTree
def tree_list_to_string(q: List[DerivationTree]) -> str:
    return "[" + ", ".join([all_terminals(tree) for tree in q]) + "]"
tree_list_to_string([derivation_tree, derivation_tree])
'[1 + (2 * 3), 1 + (2 * 3)]'

The function possible_combinations() takes a list of lists $[[x_1, x_2], [y_1, y_2], \dots]$ and creates a list of combinations $[[x_1, y_1], [x_1, y_2], [x_2, y_1], [x_2, y_2], \dots]$.

def possible_combinations(list_of_lists: List[List[Any]]) -> List[List[Any]]:
    if len(list_of_lists) == 0:
        return []

    ret = []
    for e in list_of_lists[0]:
        if len(list_of_lists) == 1:
            ret.append([e])
        else:
            for c in possible_combinations(list_of_lists[1:]):
                new_combo = [e] + c
                ret.append(new_combo)

    return ret
possible_combinations([[1, 2], ['a', 'b']])
[[1, 'a'], [1, 'b'], [2, 'a'], [2, 'b']]

The functions number_of_nodes() and max_height() return the number of nodes and the maximum height of the given tree, respectively.

def number_of_nodes(tree: DerivationTree) -> int:
    (symbol, children) = tree
    if children is None:
        return 1

    return 1 + sum([number_of_nodes(c) for c in children])
number_of_nodes(derivation_tree)
25
def max_height(tree: DerivationTree) -> int:
    (symbol, children) = tree
    if children is None or len(children) == 0:
        return 1

    return 1 + max([max_height(c) for c in children])
max_height(derivation_tree)
12

Simplification Strategies

Let us now implement our two simplification strategies – replacing subtrees and alternate expansions.

Finding Subtrees

The method subtrees_with_symbol() returns all subtrees in the given tree which's root is equal to the given symbol. If ignore_root is set (default), then the root node of tree is not compared against. (The depth parameter will be discussed below.)

class GrammarReducer(GrammarReducer):
    def subtrees_with_symbol(self, tree: DerivationTree,
                             symbol: str, depth: int = -1,
                             ignore_root: bool = True) -> List[DerivationTree]:
        """Find all subtrees in `tree` whose root is `symbol`.
        If `ignore_root` is true, ignore the root note of `tree`."""

        ret = []
        (child_symbol, children) = tree
        if depth <= 0 and not ignore_root and child_symbol == symbol:
            ret.append(tree)

        # Search across all children
        if depth != 0 and children is not None:
            for c in children:
                ret += self.subtrees_with_symbol(c,
                                                 symbol,
                                                 depth=depth - 1,
                                                 ignore_root=False)

        return ret

Here's an example: These are all subtrees with <term> in our derivation tree derivation_tree.

grammar_reducer = GrammarReducer(
    mystery,
    EarleyParser(EXPR_GRAMMAR),
    log_reduce=True)
all_terminals(derivation_tree)
'1 + (2 * 3)'
[all_terminals(t) for t in grammar_reducer.subtrees_with_symbol(
    derivation_tree, "<term>")]
['1', '(2 * 3)', '2 * 3', '3']

If we want to replace <term> subtrees to simplify the tree, these are the subtrees we could replace them with.

Alternate Expansions

Our second strategy, simplifying by alternate expansions, is a bit more complex. We first fetch the possible expansions for the given symbol (starting with the ones with the fewest children). For each expansion, we fill in values for the symbols from the subtree (using subtrees_with_symbols(), above). We then pick the first possible combination (or all combinations, if the attribute try_all_combinations is set).

class GrammarReducer(GrammarReducer):
    def alternate_reductions(self, tree: DerivationTree, symbol: str, 
                             depth: int = -1):
        reductions = []

        expansions = self.grammar.get(symbol, [])
        expansions.sort(
            key=lambda expansion: len(
                expansion_to_children(expansion)))

        for expansion in expansions:
            expansion_children = expansion_to_children(expansion)

            match = True
            new_children_reductions = []
            for (alt_symbol, _) in expansion_children:
                child_reductions = self.subtrees_with_symbol(
                    tree, alt_symbol, depth=depth)
                if len(child_reductions) == 0:
                    match = False   # Child not found; cannot apply rule
                    break

                new_children_reductions.append(child_reductions)

            if not match:
                continue  # Try next alternative

            # Use the first suitable combination
            for new_children in possible_combinations(new_children_reductions):
                new_tree = (symbol, new_children)
                if number_of_nodes(new_tree) < number_of_nodes(tree):
                    reductions.append(new_tree)
                    if not self.try_all_combinations:
                        break

        # Sort by number of nodes
        reductions.sort(key=number_of_nodes)

        return reductions
grammar_reducer = GrammarReducer(
    mystery,
    EarleyParser(EXPR_GRAMMAR),
    log_reduce=True)
all_terminals(derivation_tree)
'1 + (2 * 3)'

Here are all combinations for <term>:

grammar_reducer.try_all_combinations = True
print([all_terminals(t)
       for t in grammar_reducer.alternate_reductions(derivation_tree, "<term>")])
['1', '2', '3', '1 * 1', '1 * 3', '2 * 1', '2 * 3', '3 * 1', '3 * 3', '(2 * 3)', '1 * 2 * 3', '2 * 2 * 3', '3 * 2 * 3', '1 * (2 * 3)', '(2 * 3) * 1', '(2 * 3) * 3', '2 * (2 * 3)', '3 * (2 * 3)']

The default, though, is simply to return the first of these:

grammar_reducer.try_all_combinations = False
[all_terminals(t) for t in grammar_reducer.alternate_reductions(
    derivation_tree, "<term>")]
['1', '1 * 1']
Both Strategies Together

Let us now merge both strategies. To replace a subtree with a given symbol, we first search for already existing subtrees (using subtrees_with_symbol()); then we go for alternate expansions (using alternate_expansions()).

class GrammarReducer(GrammarReducer):
    def symbol_reductions(self, tree: DerivationTree, symbol: str, 
                          depth: int = -1):
        """Find all expansion alternatives for the given symbol"""
        reductions = (self.subtrees_with_symbol(tree, symbol, depth=depth)
                      + self.alternate_reductions(tree, symbol, depth=depth))

        # Filter duplicates
        unique_reductions = []
        for r in reductions:
            if r not in unique_reductions:
                unique_reductions.append(r)

        return unique_reductions
grammar_reducer = GrammarReducer(
    mystery,
    EarleyParser(EXPR_GRAMMAR),
    log_reduce=True)
all_terminals(derivation_tree)
'1 + (2 * 3)'

These are the possible reductions for <expr> nodes. Note how we first return subtrees (1 + (2 * 3), (2 * 3), 2 * 3) before going for alternate expansions of <expr> (1).

reductions = grammar_reducer.symbol_reductions(derivation_tree, "<expr>")
tree_list_to_string([r for r in reductions])
'[1 + (2 * 3), (2 * 3), 2 * 3, 1]'

These are the possible reductions for <term> nodes. Again, we first have subtrees of the derivation tree, followed by the alternate expansion 1 * 1.

reductions = grammar_reducer.symbol_reductions(derivation_tree, "<term>")
tree_list_to_string([r for r in reductions])
'[1, (2 * 3), 2 * 3, 3, 1 * 1]'

The Reduction Strategy

We are now able to return a number of alternatives for each symbol in the tree. This is what we apply in the core function of our reduction strategy, reduce_subtree(). Starting with subtree, for every child, we find possible reductions. For every reduction, we replace the child with the reduction and test the resulting (full) tree. If it fails, our reduction was successful; otherwise, we put the child back into place and try out the next reduction. Eventually, we apply reduce_subtree() on all children, reducing these as well.

class GrammarReducer(GrammarReducer):
    def reduce_subtree(self, tree: DerivationTree,
                       subtree: DerivationTree, depth: int = -1):
        symbol, children = subtree
        if children is None or len(children) == 0:
            return False

        if self.log_reduce:
            print("Reducing", all_terminals(subtree), "with depth", depth)

        reduced = False
        while True:
            reduced_child = False
            for i, child in enumerate(children):
                if child is None:
                    continue

                (child_symbol, _) = child
                for reduction in self.symbol_reductions(
                        child, child_symbol, depth):
                    if number_of_nodes(reduction) >= number_of_nodes(child):
                        continue

                    # Try this reduction
                    if self.log_reduce:
                        print(
                            "Replacing",
                            all_terminals(
                                children[i]),
                            "by",
                            all_terminals(reduction))
                    children[i] = reduction
                    if self.test(all_terminals(tree)) == Runner.FAIL:
                        # Success
                        if self.log_reduce:
                            print("New tree:", all_terminals(tree))
                        reduced = reduced_child = True
                        break
                    else:
                        # Didn't work out - restore
                        children[i] = child

            if not reduced_child:
                if self.log_reduce:
                    print("Tried all alternatives for", all_terminals(subtree))
                break

        # Run recursively
        for c in children:
            if self.reduce_subtree(tree, c, depth):
                reduced = True

        return reduced

All we now need is a few drivers. The method reduce_tree() is the main entry point into reduce_subtree():

class GrammarReducer(GrammarReducer):
    def reduce_tree(self, tree):
        return self.reduce_subtree(tree, tree)

The custom method parse() turns a given input into a derivation tree:

class GrammarReducer(GrammarReducer):
    def parse(self, inp):
        tree, *_ = self.parser.parse(inp)
        if self.log_reduce:
            print(all_terminals(tree))
        return tree

The method reduce() is the one single entry point, parsing the input and then reducing it.

class GrammarReducer(GrammarReducer):
    def reduce(self, inp):
        tree = self.parse(inp)
        self.reduce_tree(tree)
        return all_terminals(tree)

Let us try out our GrammarReducer class in practice on our input expr_input and the mystery() function. How quickly can we reduce it?

expr_input
'1 + (2 * 3)'
grammar_reducer = GrammarReducer(
    eval_mystery,
    EarleyParser(EXPR_GRAMMAR),
    log_test=True)
grammar_reducer.reduce(expr_input)
Test #1 '(2 * 3)' 7 FAIL
Test #2 '2 * 3' 5 PASS
Test #3 '3' 1 PASS
Test #4 '2' 1 PASS
Test #5 '(3)' 3 FAIL
'(3)'

Success! In only five steps, our GrammarReducer reduces the input to the minimum that causes the failure. Note how all tests are syntactically valid by construction, avoiding the UNRESOLVED outcomes that cause delta debugging to stall.

A Depth-Oriented Strategy

Even if five steps are already good, we can still do better. If we look at the log above, we see that after test #2, where the input (tree) is reduced to 2 * 3, our GrammarReducer first tries to replace the tree with 2 and 3, which are the alternate <term> subtrees. This may work, of course; but if there are many possible subtrees, our strategy will spend quite some time trying one after the other.

Delta debugging, as introduced above, follows the idea of trying to cut inputs approximately in half, and thus quickly proceeds towards a minimal input. By replacing a tree with much smaller subtrees, we could possibly reduce a tree significantly, but may need several attempts to do so. A better strategy is to only consider large subtrees first – both for the subtree replacement and for alternate expansions. To find such large subtrees, we limit the depth by which we search for possible replacements in the subtree – first, by looking at the direct descendants, later at lower descendants.

This is the role of the depth parameter used in subtrees_with_symbol() and passed through the invoking functions. If set, only symbols at the given depth are returned. Here's an example, starting again with our derivation tree derivation_tree:

grammar_reducer = GrammarReducer(
    mystery,
    EarleyParser(EXPR_GRAMMAR),
    log_reduce=True)
all_terminals(derivation_tree)
'1 + (2 * 3)'
display_tree(derivation_tree)
0 <start> 1 <expr> 0->1 2 <term> 1->2 7 + 1->7 8 <expr> 1->8 3 <factor> 2->3 4 <integer> 3->4 5 <digit> 4->5 6 1 (49) 5->6 9 <term> 8->9 10 <factor> 9->10 11 ( (40) 10->11 12 <expr> 10->12 24 ) (41) 10->24 13 <term> 12->13 14 <factor> 13->14 18 * 13->18 19 <term> 13->19 15 <integer> 14->15 16 <digit> 15->16 17 2 (50) 16->17 20 <factor> 19->20 21 <integer> 20->21 22 <digit> 21->22 23 3 (51) 22->23

At a depth of 1, there is no <term> symbol:

[all_terminals(t) for t in grammar_reducer.subtrees_with_symbol(
    derivation_tree, "<term>", depth=1)]
[]

At a depth of 2, we have the <term> subtree on the left-hand side:

[all_terminals(t) for t in grammar_reducer.subtrees_with_symbol(
    derivation_tree, "<term>", depth=2)]
['1']

At a depth of 3, we have the <term> subtree on the right-hand side:

[all_terminals(t) for t in grammar_reducer.subtrees_with_symbol(
    derivation_tree, "<term>", depth=3)]
['(2 * 3)']

The idea is now to start with a depth of 0, subsequently increasing it as we proceed:

class GrammarReducer(GrammarReducer):
    def reduce_tree(self, tree):
        depth = 0
        while depth < max_height(tree):
            reduced = self.reduce_subtree(tree, tree, depth)
            if reduced:
                depth = 0    # Start with new tree
            else:
                depth += 1   # Extend search for subtrees
        return tree        
grammar_reducer = GrammarReducer(
    mystery,
    EarleyParser(EXPR_GRAMMAR),
    log_test=True)
grammar_reducer.reduce(expr_input)
Test #1 '(2 * 3)' 7 FAIL
Test #2 '(3)' 3 FAIL
Test #3 '3' 1 PASS
'(3)'

We see that a depth-oriented strategy needs even fewer steps in our setting.

Comparing Strategies

We close by demonstrating the difference between text-based delta debugging and our grammar-based reduction. We build a very long expression:

from GrammarFuzzer import GrammarFuzzer
long_expr_input = GrammarFuzzer(EXPR_GRAMMAR, min_nonterminals=100).fuzz()
long_expr_input
'++---((-2 / 3 / 3 - -+1 / 5 - 2) * ++6 / +8 * 4 / 9 / 2 * 8 + ++(5) * 3 / 8 * 0 + 3 * 3 + 4 / 0 / 6 + 9) * ++++(+--9 * -3 * 7 / 4 + --(4) / 3 - 0 / 3 + 5 + 0) * (1 * 6 - 1 / 9 * 5 - 9 / 0 + 7) * ++(8 - 1) * +1 * 7 * 0 + ((1 + 4) / 4 * 8 * 9 * 4 + 4 / (4) * 1 - (4) * 8 * 5 + 1 + 4) / (+(2 - 1 - 9) * 5 + 3 + 6 - 2) * +3 * (3 - 7 + 8) / 4 - -(9 * 4 - 1 * 0 + 5) / (5 / 9 * 5 + 2) * 7 + ((7 - 5 + 3) / 1 * 8 - 8 - 9) * --+1 * 4 / 4 - 4 / 7 * 4 - 3 / 6 * 1 - 2 - 7 - 8'

With grammars, we need only a handful of tests to find the failure-inducing input:

from Timer import Timer
grammar_reducer = GrammarReducer(eval_mystery, EarleyParser(EXPR_GRAMMAR))
with Timer() as grammar_time:
    print(grammar_reducer.reduce(long_expr_input))
(9)
grammar_reducer.tests
10
grammar_time.elapsed_time()
0.0741315829945961

Delta debugging, in contrast, requires orders of magnitude more tests (and consequently, time). Again, the reduction is not closely as perfect as it is with the grammar-based reducer.

dd_reducer = DeltaDebuggingReducer(eval_mystery)
with Timer() as dd_time:
    print(dd_reducer.reduce(long_expr_input))
((2 - 1 - 2) * 8 + (5) - (4)) / ((2) * 3) * (9) / 3 / 1 - 8
dd_reducer.tests
900
dd_time.elapsed_time()
2.5649758750078036

We see that if an input is syntactically complex, using a grammar to reduce inputs is the best way to go.

Lessons Learned

  • Reducing failure-inducing inputs to a minimum is helpful for testing and debugging.
  • Delta debugging is a simple and robust algorithm to easily reduce test cases.
  • For syntactically complex inputs, grammar-based reduction is much faster and yields better results.

Next Steps

Our next chapter focuses on Web GUI Fuzzing, another domain where generating and reducing test cases is central.

Background

The "lexical" delta debugging algorithm discussed here stems from [Zeller et al, 2002]; actually, this is the exact Python implementation as used by Zeller in 2002. The idea of systematically reducing inputs has been discovered a number of times, although not as automatic and generic as delta debugging. [Slutz et al, 1998], for instance, discusses systematic reduction of SQL statements for SQL databases; the general process as manual work is well described by [Kernighan et al, 1999].

The deficits of delta debugging as it comes to syntactically complex inputs were first discussed in compiler testing, and reducing tree inputs rather than string inputs was quickly discovered as an alternative. Hierarchical Delta Debugging (HDD) [Misherghi et al, 2006] applies delta debugging on subtrees of a parse tree, systematically reducing a parse tree to a minimum. Generalized Tree Reduction [Herfert et al, 2017] generalizes this idea to apply arbitrary patterns such as replacing a term by a compatible term in a subtree, as subtrees_with_symbol() does. Using grammars to reduce inputs was first implemented in the Perses tool [Sun et al, 2018]; our algorithm implements very similar strategies. Searching for alternate expansions (as alternate_reductions()) is a contribution of the present chapter.

While applying delta debugging to code lines does a decent job, syntactic and especially language-specific approaches can do a much better job for the programming language at hand:

  • C-Reduce [Regehr et al, 2012] is a reducer specifically targeting the reduction of programming languages. Besides reductions in the style of delta debugging or tree transformations, C-Reduce comes with more than 30 source-to-source transformations that replace aggregates by scalars, remove function parameters at a definition and all call sites, change functions to return void and deleting all return statements, and many more. While specifically instantiated for the C language (and used for testing C compilers), these principles extend to arbitrary programming languages following an ALGOL-like syntax.

  • Kalhauge and Palsberg [Kalhauge et al, 2019] introduce binary reduction of dependency graphs, a general solution for reducing arbitrary inputs with dependencies. Their J-Reduce tool specifically targets Java programs, and again is much faster than delta debugging and achieves a higher reduction rate.

Reducing inputs also works well in the context of property-based testing; that is, generating data structures for individual functions, which can then be reduced ("shrunk") upon failure. The Hypothesis fuzzer has a number of type-specific shrinking strategies; this blog article discusses some of its features.

The chapter on "Reducing Failure-Inducing Inputs" in the Debugging Book has an alternate implementation DeltaDebugger of delta debugging that is even easier to deploy; here, one simply writes

with DeltaDebugger() as dd:
    fun(args...)
dd

to reduce the input in args for a failing (exception-throwing) function fun(). The chapter also discusses further usage examples, including reducing code to a minimum.

This blog post by David McIver contains lots of insights on how to apply reduction in practice, in particular multiple runs with different abstraction levels.

Exercises

How to best reduce inputs is still an underdeveloped field of research, with lots of opportunities.

Exercise 1: Mutation-Based Fuzzing with Reduction

When fuzzing with a population, it can be useful to occasionally reduce the length of each element, such that future descendants are shorter, too, which typically speeds up their testing.

Consider the MutationFuzzer class from the chapter on mutation-based fuzzing. Extend it such that whenever a new input is added to the population, it is first reduced using delta debugging.

Exercise 2: Reduction by Production

Grammar-based input reduction, as sketched above, might be a good algorithm, but is by no means the only alternative. One interesting question is whether "reduction" should only be limited to elements already present, or whether one would be allowed to also create new elements. These would not be present in the original input, yet still allow producing a much smaller input that would still reproduce the original failure.

As an example, consider the following grammar:

<number> ::= <float> | <integer> | <not-a-number>
<float> ::= <digits>.<digits>
<integer> ::= <digits>
<not-a-number> ::= NaN
<digits> ::= [0-9]+

Assume the input 100.99 fails. We might be able to reduce it to a minimum of, say, 1.9. However, we cannot reduce it to an <integer> or to <not-a-number>, as these symbols do not occur in the original input. By allowing to create alternatives for these symbols, we could also test inputs such as 1 or NaN and further generalize the class of inputs for which the program fails.

Create a class GenerativeGrammarReducer as subclass of GrammarReducer; extend the method reduce_subtree() accordingly.

Exercise 3: The Big Reduction Shoot-Out

Create a benchmark for the grammars already defined earlier, consisting of:

  1. A set of inputs, produced from these very grammars using GrammarFuzzer and derivatives;
  2. A set of tests which check for the occurrence of individual symbols as well as pairs and triples of these symbols:
    • Tests should be unresolved if the input is not syntactically valid;
    • Tests should fail if the symbols (or pairs or triples thereof) occur;
    • Tests should pass in all other cases.

Compare delta debugging and grammar-based debugging on the benchmark. Implement HDD [Misherghi et al, 2006] and Generalized Tree Reduction [Herfert et al, 2017] and add them to your comparison. Which approach performs best, and under which circumstances?

Creative Commons License The content of this project is licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License. The source code that is part of the content, as well as the source code used to format and display that content is licensed under the MIT License. Last change: 2023-11-11 18:18:06+01:00CiteImprint